An information flow recommendation method, system and medium based on a pre-training algorithm

Abstract

The application discloses a kind of information stream recommendation method, system and medium based on pre-training algorithm, the method includes the following steps: based on big data platform, pre-processing parameter and first training algorithm call pre-training algorithm, obtain pre-training word vector set;User historical operation behavior data is obtained, based on vector retrieval model and user historical operation behavior data call information recall algorithm, obtain recall information set;Based on pre-training word vector set, recall information set, first feature acquisition algorithm, second feature acquisition algorithm, similarity analysis algorithm, deep learning model and fine arrangement model execute information sorting recommendation operation;The application can be based on self-research algorithm to carry out the accurate matching recommendation of information, and the range of recommended information matching is low in limitation, easy to expand and extend, in numerous platform information, algorithm can be called according to user demand from multiple levels to carry out information screening interception, adds innovative sorting feature in fine arrangement stage, improves recommended accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information recommendation, in particular to an information stream recommendation method and system based on a pre-training algorithm and a medium. BACKGROUND

[0002] With the rapid development of the Internet, information overload has become a common phenomenon, and each information service platform is focusing on how to efficiently provide high-quality information to users to improve their experience. In order to deal with the above situation, information recommendation has become an effective solution.

[0003] At present, information recommendation methods focus on content recommendation and commodity recommendation, and information recommendation methods are mainly based on information filtering "recall" and information filtering "ranking" to achieve. In the traditional "recall" mode, the user preference label is calculated based on the statistical strategy, and then the information is recalled according to the user preference label. This recall method has certain homogeneity for users, and the diversity is low, and the recalled information is also some high-hot information for users. With the long-term development of the platform, this information recall method will make the recommendation range of information narrower and narrower, which is not conducive to the promotion and distribution of other business content of the platform.

[0004] In summary, it is necessary to develop an information stream recommendation method with high diversity of recommendation range and at the same time meeting the personalized needs of users, and high accuracy, so as to meet the needs of users while meeting the fairness and diversity of platform content promotion and distribution. SUMMARY

[0005] The main purpose of the present application is to develop an information stream recommendation method with high diversity of recommendation range and at the same time meeting the personalized needs of users, and high accuracy, so as to meet the needs of users while meeting the fairness and diversity of platform content promotion and distribution.

[0006] To achieve the above purpose, one technical scheme adopted by the present application is to provide an information stream recommendation method based on a pre-training algorithm, comprising the following steps:

[0007] Pre-training processing step:

[0008] Set the pre-processing parameters, configure the pre-training algorithm and the first training algorithm, call the pre-training algorithm based on the big data platform, the pre-processing parameters and the first training algorithm, and obtain the pre-training word vector set;

[0009] Recall pool configuration step:

[0010] Configure the vector retrieval model and the information recall algorithm, obtain the user historical operation behavior data, call the information recall algorithm based on the vector retrieval model and the user historical operation behavior data, and obtain the recall information set.

[0011] The sorting processing step includes:

[0012] The first feature acquisition algorithm, the second feature acquisition algorithm, the similarity analysis algorithm, the deep learning model and the fine ranking model are configured, and information sorting recommendation operations are performed based on the pre-trained word vector set, the recall information set, the first feature acquisition algorithm, the second feature acquisition algorithm, the similarity analysis algorithm, the deep learning model and the fine ranking model.

[0013] As an improved scheme, the preprocessing parameters include: a first time range, a second time range and a third training period;

[0014] The pre-training algorithm is:

[0015] Based on the big data platform, first behavior data of the user within the first time range is obtained; information click sequences of the user are obtained based on the first behavior data; the information click sequences are taken as training data, and the first training algorithm is called to train the training data according to the third training period, to obtain a plurality of pre-training information word vectors;

[0016] A plurality of first word vectors are confirmed from a plurality of pre-training information word vectors according to the first time range; a plurality of second word vectors are confirmed from a plurality of pre-training information word vectors according to the second time range;

[0017] Based on a plurality of first word vectors, a long-term pre-training user word vector is obtained by weighted average calculation; based on a plurality of second word vectors, a short-term pre-training user word vector is obtained by weighted average calculation;

[0018] The long-term pre-training user word vector, the short-term pre-training user word vector and a plurality of pre-training information word vectors are packaged to obtain the pre-trained word vector set.

[0019] As an improved scheme, the information recall algorithm is:

[0020] A similarity reference value is set, a similarity reference recall step is performed based on the vector retrieval model, the user historical operation behavior data and the similarity reference value, and first recall information is obtained;

[0021] A weight reference value is set, a weight value reference recall step is performed based on the user historical operation behavior data and the weight reference value, and second recall information is obtained;

[0022] A click frequency threshold and a probability reference value are set, a probability value reference recall step is performed based on the user historical operation behavior data, the click frequency threshold and the probability reference value, and third recall information is obtained;

[0023] packaging the first recall information, the second recall information and the third recall information to obtain the recall information set.

[0024] As an improved scheme, the similarity reference recall step comprises:

[0025] Based on the user historical operation behavior data, user interest information is identified; a first interest word vector matching the user interest information is screened from a plurality of pre-training information word vectors; a first similar vector with a vector similarity reaching the similarity reference value with the first interest word vector is confirmed based on the vector retrieval model; first similar information matching the first similar vector is intercepted from the platform information library to obtain the first recall information.

[0026] As an improved scheme, the weight value reference recall step comprises:

[0027] Based on the user historical operation behavior data, user preference data is calculated; preference information matching the user preference data is screened from the platform information library; a preference weight of the preference information is calculated based on the user preference data; information quality of the preference information is identified; an information weight of the preference information is calculated based on the preference weight and the information quality; the preference information corresponding to the information weight reaching the weight reference value is intercepted as the second recall information.

[0028] As an improved scheme, the probability value reference recall step comprises:

[0029] Based on the user historical operation behavior data, user information click data is identified; high-frequency click information reaching the click frequency threshold is identified in the user historical operation behavior data based on the user information click data; a secondary click probability of the high-frequency click information is predicted based on Poisson distribution; the high-frequency click information corresponding to the secondary click probability reaching the probability reference value is intercepted as the third recall information.

[0030] As an improved scheme, the information sorting recommendation operation comprises:

[0031] The first feature acquisition algorithm is called to obtain user coarse ranking features and information coarse ranking features.

[0032] The similarity analysis algorithm is called based on the pre-training word vector set, the recall information set, the user coarse ranking features, the information coarse ranking features and the deep learning model to obtain a coarse ranking information set.

[0033] The second feature acquisition algorithm is called to acquire fine arrangement features; the fine arrangement model is trained based on the fine arrangement features to obtain a to-be-used ranking model; the to-be-used ranking model is called to perform fine arrangement and screening processing on the coarse arrangement information set to obtain a fine arrangement information set.

[0034] The fine arrangement information set is output to a user end.

[0035] As an improved scheme, the similarity analysis algorithm is:

[0036] A first similarity index is set;

[0037] The user coarse arrangement feature is introduced into an embedding layer of the deep learning model for calculation to obtain a user coarse arrangement feature word vector; the pre-trained word vector set is accessed, and the user coarse arrangement feature word vector, the short-term pre-trained user word vector and the long-term pre-trained user word vector are connected to obtain a user mixed word vector;

[0038] The information coarse arrangement feature is introduced into the embedding layer of the deep learning model for calculation to obtain an information coarse arrangement feature word vector; the pre-trained word vector set is accessed, and the information coarse arrangement feature word vector and a plurality of pre-trained information word vectors are connected respectively to obtain a plurality of information mixed word vectors;

[0039] The user mixed word vector and a plurality of information mixed word vectors are respectively subjected to similarity calculation based on a cosine formula by entering a full connection layer of the deep learning model to obtain a plurality of similarity results;

[0040] The similarity result reaching the first similarity index is set as a qualified result; and the information mixed word vector corresponding to the qualified result is set as a qualified vector;

[0041] The recall information set is intercepted to obtain the coarse arrangement information set.

[0042] The application further provides an information flow recommendation system based on a pre-training algorithm, comprising:

[0043] A pre-training processing module, a recall pool configuration module and a ranking processing module;

[0044] The pre-training processing module is used for setting pre-processing parameters, configuring a pre-training algorithm and a first training algorithm, calling the pre-training algorithm based on a big data platform, the pre-processing parameters and the first training algorithm to obtain a pre-trained word vector set;

[0045] The recall pool configuration module is configured to configure a vector retrieval model and an information recall algorithm, and obtain user historical operation behavior data, and the recall pool configuration module calls the information recall algorithm based on the vector retrieval model and the user historical operation behavior data to obtain a recall information set.

[0046] The sorting processing module is configured to configure a first feature acquisition algorithm, a second feature acquisition algorithm, a similarity analysis algorithm, a deep learning model and a fine ranking model, and the sorting processing module performs information sorting and recommendation operations based on the pre-trained word vector set, the recall information set, the first feature acquisition algorithm, the second feature acquisition algorithm, the similarity analysis algorithm, the deep learning model and the fine ranking model.

[0047] The application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the information stream recommendation method based on a pre-trained algorithm.

[0048] The application has the following advantages:

[0049] 1、The information stream recommendation method based on a pre-trained algorithm can realize accurate matching and recommendation of platform information based on self-developed algorithms, has low range limitation of recommended information matching, is easy to expand and extend, can call algorithms to perform information filtering and interception from multiple aspects according to user needs in numerous information on the platform, and further improves the accuracy and matching degree of recommended information by adding self-innovative sorting features in the final fine ranking stage, thereby improving user experience and information recommendation processing efficiency, and having high application value and market value.

[0050] 2、The information stream recommendation system based on a pre-trained algorithm can realize accurate matching and recommendation of platform information based on self-developed algorithms through mutual cooperation of the pre-training processing module, the recall pool configuration module and the sorting processing module, has low range limitation of recommended information matching, is easy to expand and extend, can call algorithms to perform information filtering and interception from multiple aspects according to user needs in numerous information on the platform, and further improves the accuracy and matching degree of recommended information by adding self-innovative sorting features in the final fine ranking stage, thereby improving user experience and information recommendation processing efficiency, and having high application value and market value.

[0051] 3. The computer-readable storage medium described in this invention enables the pre-training processing module, the recall pool configuration module, and the sorting processing module to work together to achieve accurate matching and recommendation of platform information based on a self-developed algorithm. Furthermore, the scope of recommended information matching is minimal, and it is easily expanded and extended. Among the vast amount of information on the platform, the algorithm can be invoked to filter and extract information from multiple levels according to user needs. In the final fine-ranking stage, an independently developed sorting feature is added, further improving the accuracy and matching degree of the recommended information. Ultimately, this not only enhances the user experience but also improves the efficiency of information recommendation processing and effectively enhances the operability of the information flow recommendation method based on the pre-training algorithm. Attached Figure Description

[0052] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the information flow recommendation method based on pre-trained algorithms described in Embodiment 1 of the present invention;

[0054] Figure 2 This is a schematic diagram of the specific process of the information flow recommendation method based on pre-trained algorithm described in Embodiment 1 of the present invention;

[0055] Figure 3 This is an architecture diagram of the information flow recommendation system based on pre-trained algorithms described in Embodiment 2 of the present invention. Detailed Implementation

[0056] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0057] In the description of this invention, it should be noted that the embodiments described in this invention are only some embodiments of this invention, not all embodiments; based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0058] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0059] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "preprocessing parameter", "pre-training algorithm", "pre-training word vector set", "vector retrieval model", "information recall algorithm", "user historical operation behavior data", "recalled information set", "feature acquisition algorithm", "similarity analysis algorithm", "deep learning model", "fine ranking model", "information ranking recommendation operation", "time range", "training period", "behavior data", "information click sequence", "training data", "pre-training information word vector", "weighted average calculation", "long-term pre-training user word vector", "short-term pre-training user word vector", "similarity benchmark value", "weight benchmark value", "click frequency threshold", "probability benchmark value", "user interest information", "vector similarity", "user preference data", "preference information", "preference weight", "information quality", "information weight", "user information click data", "high-frequency click information", "secondary click probability", "user coarse ranking feature", "information coarse ranking feature", "to-be-used ranking model", "fine ranking screening processing", "similarity index", "user coarse ranking feature word vector", "user mixed word vector", "information coarse ranking feature word vector", "information mixed word vector", "cosine formula", "similarity calculation", "similarity result", "standard reaching result", "standard reaching vector", "pre-training processing module", "recall pool configuration module", "ranking processing module" should be understood broadly. For ordinary skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0060] In the description of the present application, it should be noted that: CTR (Click-Through-Rate) is the click-through rate.

[0061] Embodiment 1

[0062] The present embodiment provides an information stream recommendation method based on a pre-training algorithm, as shown in Figure 1 and Figure 2 , comprising the following steps:

[0063] S100, pre-training processing step, specifically comprising:

[0064] S110, set a preprocessing parameter, configure a pre-training algorithm and a first training algorithm, call the pre-training algorithm based on a big data platform, the preprocessing parameter and the first training algorithm, and obtain a pre-training word vector set; in this example, the main purpose of step S100 is to perform vectorization calculation of information according to the behavior sequence of the user before the recall, rough sorting and fine sorting steps of information recommendation, and then pre-train a plurality of vectors representing the user and the information as a data basis, so as to perform high-precision and low-limit information recommendation matching based on the data basis in the subsequent information recommendation process; correspondingly, in this embodiment, the article recommendation applied to the data platform is taken as an example of an embodiment of the method.

[0065] Specifically, the preprocessing parameter includes a first time range, a second time range and a third training period; in this embodiment, the first time range is within 30 days before the time of the last login of the user to the platform, the second time range is within 5 days before the time of the last login of the user to the platform, and the third training period is set to 1 hour in this embodiment.

[0066] Specifically, the pre-training algorithm is: first, the big data platform collects the first behavior data of the user to be recommended in the first time range, that is, the user's behavior event data in the platform in the past 30 days, and the behavior time data is the first behavior data, including the user's article click behavior, link click behavior, video click behavior, picture click behavior, browsing behavior of the above content, commodity purchase behavior and commodity sharing behavior, etc.; based on the first behavior data, the information click sequence of the user is obtained, that is, the information sequence of the article clicked by the user extracted from the first behavior data, which contains all the articles clicked by the user in the past 30 days; the information click sequence is taken as the training data, and the first training algorithm is called according to the third training period to train the training data, and a plurality of pre-training information word vectors are obtained; in this embodiment, the first training algorithm is item2vec / node2vec algorithm, and the training data is trained every hour according to the third training period, and the finally generated a plurality of pre-training information word vectors are the word vectors of each article in the information click sequence respectively, which can be used as the representative vector of the article; the pre-training information word vector is the article word vector generated by the initial pre-training of this method, and the user word vector also needs to be pre-trained subsequently; the user word vector is a word vector that can represent the user, so a plurality of first word vectors are confirmed from a plurality of pre-training information word vectors according to the first time range; a plurality of second word vectors are confirmed from a plurality of pre-training information word vectors according to the second time range; correspondingly, in this embodiment, the above steps are to confirm all word vectors of the articles clicked by the user in the past 30 days, that is, a plurality of first word vectors, and to confirm all word vectors of the articles clicked by the user in the past 5 days, that is, a plurality of second word vectors, in a plurality of pre-training information word vectors; based on the first time range and the second time range, the obtained word vectors are weighted and averaged respectively to obtain long-term word vectors and short-term word vectors that can represent the user; therefore, the long-term pre-training user word vector is obtained by weighted average calculation based on a plurality of first word vectors; the short-term pre-training user word vector is obtained by weighted average calculation based on a plurality of second word vectors; the finally obtained word vector can be stored in the big data platform for subsequent use, so the long-term pre-training user word vector, the short-term pre-training user word vector and a plurality of pre-training information word vectors are packaged and stored to obtain the pre-training word vector set.

[0067] S200, recall pool configuration step, specifically comprising:

[0068] S210, configure a vector retrieval model and an information recall algorithm, obtain user historical operation behavior data, call the information recall algorithm based on the vector retrieval model and the user historical operation behavior data, and obtain a recall information set; in this embodiment, step S200 is to use the self-developed algorithm in the method to recall high-relevance articles related to the user as a data pool for subsequent article recommendation for the user.

[0069] Specifically, the information recall algorithm is: setting a similarity reference value, performing a similarity reference recall step based on the vector retrieval model, the user historical operation behavior data and the similarity reference value to obtain first recall information; in this embodiment, the similarity reference value is set according to specific circumstances, which represents the similarity between vectors. The vector similarity reaching the similarity reference value is high similarity, and the vector similarity not reaching the similarity reference value is low similarity; setting a weight reference value, performing a weight value reference recall step based on the user historical operation behavior data and the weight reference value to obtain second recall information; in this embodiment, the weight reference value is set according to specific circumstances, which is used as a weight differentiation reference. The weight value reaching the weight reference value is judged as high weight, and the weight value not reaching the weight reference value is judged as low weight; setting a click frequency threshold and a probability reference value, performing a probability value reference recall step based on the user historical operation behavior data, the click frequency threshold and the probability reference value to obtain third recall information; in this embodiment, the click frequency threshold and the probability reference value are both set according to specific circumstances. The click frequency threshold is used to judge the high-frequency clicked articles of the user. If the click frequency reaches the click frequency threshold, it means that the article corresponding to the click frequency is a high-frequency clicked article of the user. The probability reference value is used to judge the secondary click probability of the article. If the secondary click probability of the article reaches the probability reference value, it means that the secondary click probability of the article is high, otherwise it is low. Correspondingly, the similarity reference recall step, the weight value reference recall step and the probability value reference recall step are to filter out articles with high interest and high relevance matching the user as data basis in the user historical operation behavior data according to different reference values. According to different reference values, the diversity and accuracy of the method can be further improved. Therefore, the first recall information, the second recall information and the third recall information are packaged to obtain the recall information set.

[0070] Specifically, the similarity reference recall step includes: identifying user interest information based on the user historical operation behavior data; in this embodiment, the user historical operation behavior data includes but is not limited to any log data of user historical clicks, collections, likes, comments or purchases; the user interest information is the article that the user is interested in; in this embodiment, the way of obtaining the article that the user is interested in is not limited, for example, the user historical operation behavior data can be divided into the same category, then the category and the subcategory with the largest data quantity ratio are the categories that the user is interested in, and the articles under the corresponding categories are the articles that the user is interested in; the user historical operation behavior data in this step and the first behavior data in step S100 belong to the same log data range of the user; therefore, the user interest information is used as the distinguishing identifier, a first interest word vector that matches the user interest information is selected from the plurality of pre-training information word vectors, the first interest word vector is the word vector corresponding to the article that the user is interested in; therefore, based on the vector retrieval model, a first similar vector with a vector similarity to the first interest word vector reaching the similarity reference value is confirmed, the first similar vector reaching the similarity reference value is a word vector with high similarity to the word vector corresponding to the article that the user is interested in, and the article or information corresponding to the first similar vector can be an article with high relevance or high interest degree to the user, which can be recalled; therefore, the first similar information matching the first similar vector is intercepted in the platform information library to obtain the first recall information; correspondingly, in this embodiment, the vector retrieval model includes but is not limited to an Item2Vec model and a DeepWalk model, the above models can be used to train the user behavior sequence in a certain way, and the model that generates the vectorization result of the corresponding user behavior sequence; correspondingly, the vector retrieval model in this embodiment includes but is not limited to the above models and other models that can achieve the same function; the platform information library is a database in which all articles of the platform to which this method is applied are stored; and the vector similarity between word vectors is calculated based on the spatial distance between word vectors.

[0071] Specifically, the weight value reference recall step includes: based on the user historical operation behavior data, the user portrait is depicted, and the user preference data is calculated according to the user portrait depiction; correspondingly, because the method is applied to the vehicle information platform, therefore, the user preference data usually includes but is not limited to vehicle type preference, vehicle brand preference and vehicle accessory preference, etc.; therefore, the preference information matched with the user preference data is screened out in the platform information database, and the preference information is all the articles corresponding to the user preference data; the preference weight of the preference information is calculated based on the user preference data, and correspondingly, in the embodiment, the calculation of the preference weight is as follows: the preference weight of a certain preference information = the number of times that the user clicks the preference information in the same time period / the total number of times that the user clicks, for example, the preference weight of a certain preference information = the number of times that the user clicks the preference information in the recent N days / the total number of times that the user clicks any information in the recent N days; the information quality of the preference information is identified, and correspondingly, in the embodiment, the information quality is the quality of the corresponding preference article, and the calculation logic is: the click conversion rate of the article, the title length of the article, the content length of the article and the number of pictures of the article are weighted and summed to calculate, and the specific formula is: information quality = 0.5*CTR + 0.2*title length + 0.2*content length + 0.1*picture number; therefore, the preference weight and the information quality of the preference information corresponding to the preference weight are taken as two elements to calculate the information weight of the preference information, and finally the information weight of each preference information is obtained; the preference information corresponding to the information weight reaching the weight reference value is intercepted as the second recall information, and the preference information corresponding to the information weight reaching the weight reference value is the user preference article with high user correlation degree and high article quality, and through the calculation logic in this step, most of the low-quality articles can be screened out to ensure the high effectiveness of the articles in the recall data pool.

[0072] Specifically, the probability value reference recall step includes: identifying user information click data based on the user historical operation behavior data, the user information click data being the number of times that a user clicks a specific article and the article information corresponding to the number of times; identifying high-frequency click information that reaches the click number threshold in a unit time in the user historical operation behavior data based on the user information click data, the high-frequency click information that reaches the click number threshold being judged as article information that meets the Poisson distribution; predicting the secondary click probability of the high-frequency click information based on the Poisson distribution, the secondary click probability being the probability that the user clicks the high-frequency click information again; the high-frequency click information corresponding to the secondary click probability that reaches the probability reference value is information with high relevance to the user, so the high-frequency click information corresponding to the secondary click probability that reaches the probability reference value is intercepted as the third recall information; correspondingly, the final first recall information, second recall information and third recall information in this embodiment are all in the form of articles, and the final recall information set is a recall data pool composed of a plurality of articles.

[0073] S300, an ordering processing step, specifically includes:

[0074] S310, configuring a first feature acquisition algorithm, a second feature acquisition algorithm, a similarity analysis algorithm, a deep learning model and a fine ranking model, performing information ordering recommendation operation based on the pre-trained word vector set, the recall information set, the first feature acquisition algorithm, the second feature acquisition algorithm, the similarity analysis algorithm, the deep learning model and the fine ranking model; correspondingly, in step S300, the pre-trained word vector obtained in step S100 is needed to perform re-selection on the recall information set obtained in step S200, and the articles with higher user interest degree are ranked to high positions, the main logic steps are divided into coarse ranking and fine ranking, coarse ranking is performed first and fine ranking is performed last, and in the coarse ranking and fine ranking processes, unique algorithm logic in the method is adopted;

[0075] Specifically, the information ranking recommendation operation includes: calling the first feature acquisition algorithm to obtain user coarse ranking features and information coarse ranking features; in this embodiment, the first feature acquisition algorithm is based on data analysis and mining technology, and the features of the user side and the article side are extracted to obtain the user coarse ranking features and the information coarse ranking features; correspondingly, in this embodiment, the user coarse ranking features include but are not limited to user attribute features, user behavior features, and user statistical features, etc., the user attribute features include but are not limited to user gender, user age, user occupation, and user region, etc., the user behavior features include but are not limited to user browsing time length for different time windows, user click behavior for different time windows, user search behavior for different time windows, user posting behavior for different time windows, user collection behavior for different time windows, and user like behavior for different time windows, etc., and the user statistical features include but are not limited to user registration time, user activity, and user cumulative online time, etc. statistical index features; correspondingly, the information coarse ranking features in this embodiment include but are not limited to article title word length, article word total length, article browsing time length, article click number, article exposure number, and article click rate, etc.; therefore, based on the pre-trained word vector set, the recall information set, the user coarse ranking features, the information coarse ranking features, and the deep learning model, the similarity analysis algorithm is called to obtain a coarse ranking information set; the coarse ranking information set is a data pool after one round of screening of the articles in the recall information set by the similarity analysis algorithm, and the articles have higher interest accuracy and relevance to the user than the recall information set; then the second feature acquisition algorithm is called to obtain fine ranking features; the fine ranking model is trained based on the fine ranking features to obtain a to-be-used ranking model; the coarse ranking information set is screened and processed by the to-be-used ranking model to obtain a fine ranking information set; and the fine ranking information set is output to the user end;

[0076] Specifically, the similarity analysis algorithm is: a first similarity index is set, in this embodiment, the first similarity index is used to further judge the relevance between the information word vector and the user word vector; therefore, the user coarse ranking feature is introduced into the embedding layer of the deep learning model for calculation to obtain a user coarse ranking feature word vector; the pre-trained word vector set is accessed, the user coarse ranking feature word vector, the short-term pre-trained user word vector and the long-term pre-trained user word vector are connected for processing to obtain a user mixed word vector; the information coarse ranking feature is introduced into the embedding layer of the deep learning model for calculation to obtain an information coarse ranking feature word vector; the pre-trained word vector set is accessed, the information coarse ranking feature word vector and each pre-trained information word vector in the pre-trained information word vector are connected for processing to obtain a plurality of information mixed word vectors; the connection processing is to connect several word vectors; after the above connection is completed, the full connection layer of the deep learning model is entered, the user mixed word vector and the plurality of information mixed word vectors are respectively calculated based on the cosine formula to obtain a plurality of similarity results; the similarity result reaching the first similarity index is set as a qualified result, the information (in this embodiment, it is an article) corresponding to the information mixed word vector corresponding to the qualified result is information with high relevance to the user interest; therefore, the information mixed word vector corresponding to the qualified result is set as a qualified vector; the recall information matching the qualified vector in the recall information set is packaged to obtain the coarse ranking information set; the coarse ranking information set is further filtered based on the similarity analysis algorithm on the recall data set, and the matching degree between the article and the user is improved; correspondingly, in this embodiment, when applied, the obtained qualified vector needs to be stored in the clustering index database for saving, so as to facilitate the subsequent use.

[0077] Specifically, in the present embodiment, the second feature acquisition algorithm is to set innovative features matching the application platform type and acquire the innovative features. Since the present embodiment is applied to a vehicle information platform, the innovative features include but are not limited to whether the article is associated with a vehicle, the vehicle brand corresponding to the article, and the vehicle category corresponding to the article. Correspondingly, the innovative features are the above-mentioned fine arrangement features. In the present embodiment, for the acquisition of the "whether the article is associated with a vehicle" feature, the article keywords are extracted by the TF-IDF and TextRank methods, and the keywords extracted by the two methods are fused and input into a binary classification model for probability calculation, and the keywords with high probability are selected as the corresponding feature data. The "vehicle brand corresponding to the article" feature is acquired by using a sequence labeling method. The specific principle is to train a sequence labeling model using article samples labeled with relevant vehicle brands, and to acquire vehicle brand related data of unlabeled articles using the sequence labeling model. Correspondingly, the sequence labeling model structure is BiLSTM+CRF structure. The "vehicle category corresponding to the article" feature is acquired by using a multi-classification model to predict the vehicle category data of articles with unlabeled vehicle categories in the platform. The multi-classification model is a model trained using known vehicle category data in the platform and articles labeled with vehicle category data in the platform. Correspondingly, the model is TextCNN. Finally, the above-mentioned innovative features obtain the feature data of articles unrelated to vehicles in the application platform, as well as all vehicle category and vehicle brand feature data of articles related to vehicles. After obtaining the above-mentioned data, the fine arrangement processing accuracy of the fine arrangement model can be further improved. Therefore, training the fine arrangement model based on the features is to add the above-mentioned innovative features to the ranking features of the fine arrangement model to train the fine arrangement model, so that the fine arrangement model learns whether the user reads the article associated with the vehicle and whether the user reads the article associated with the corresponding vehicle brand and vehicle category. First, the training set is sampled based on the above-mentioned innovative features, and then the training and testing are performed to obtain the to-be-used ranking model that can be applied. Finally, when applied, the to-be-used ranking model only needs to be embedded in the recommendation system of the platform to which the present method is applied for fine arrangement processing. Finally, the to-be-used ranking model performs fine arrangement and screening processing on the coarse arrangement information set, and the fine arrangement information set obtained is highly relevant to the user and matches the user's reading habits. There are two cases, one is that the user's reading habits are not associated with vehicles, and then the to-be-used ranking model will automatically identify and screen out the articles unrelated to vehicles in the coarse arrangement information set, so that the fine arrangement information set will contain the articles unrelated to vehicles and interesting to the user in the coarse arrangement information set. The other is that the user's reading habits are associated with vehicles, and then the to-be-used ranking model will screen out the articles unrelated to vehicles in the coarse arrangement information set, so that the fine arrangement information set will contain the articles related to vehicles and specific to the corresponding vehicle brand and vehicle category in the coarse arrangement information set.Correspondingly, in this step, when the refined information set is output to the user end, the information quality weight of the article in the refined information set also needs to be calculated based on the calculation logic in the previous step S200, and then the corresponding refined article sequence is generated according to the weight from large to small, and finally the refined article sequence is output to the user end, so that the article most interested by the user is located at the front of the sequence.

[0078] Through effective implementation calculation, the articles and information recommended by the method have obvious improvement compared with traditional recommendation methods, and the effective reading time of the user for the articles recommended by the method is obviously improved. The recall, rough sorting and refined sorting range of the method are not limited, and the diversity of the platform to which the method is applied will not be reduced, which makes up for the shortcomings of the prior art.

[0079] Embodiment 2

[0080] This embodiment is based on the same inventive concept as the information flow recommendation method based on pre-training algorithm described in embodiment 1, and provides an information flow recommendation system based on pre-training algorithm, as shown in Figure 3 The information flow recommendation system based on pre-training algorithm includes a pre-training processing module, a recall pool configuration module and a sorting processing module.

[0081] In the information flow recommendation system based on pre-training algorithm, the pre-training processing module is used to set pre-processing parameters and configure pre-training algorithms and first training algorithms. The pre-training processing module calls the pre-training algorithm based on a big data platform, the pre-processing parameters and the first training algorithms to obtain a pre-training word vector set.

[0082] Specifically, the pre-processing parameters include a first time range, a second time range and a third training period.

[0083] Specifically, the pre-training algorithm is: the pre-training processing module acquires first behavior data of a user in the first time range based on the big data platform; the pre-training processing module acquires an information click sequence of the user based on the first behavior data; the pre-training processing module takes the information click sequence as training data, and trains the training data according to the first training algorithm in the third training period to obtain a plurality of pre-training information word vectors; the pre-training processing module confirms a plurality of first word vectors in the plurality of pre-training information word vectors according to the first time range; the pre-training processing module confirms a plurality of second word vectors in the plurality of pre-training information word vectors according to the second time range; the pre-training processing module performs weighted average calculation based on the plurality of first word vectors to obtain a long-term pre-training user word vector; the pre-training processing module performs weighted average calculation based on the plurality of second word vectors to obtain a short-term pre-training user word vector; and the pre-training processing module packages the long-term pre-training user word vector, the short-term pre-training user word vector, and the plurality of pre-training information word vectors to obtain the pre-training word vector set.

[0084] In the information flow recommendation system based on the pre-training algorithm, the recall pool configuration module is configured to configure a vector retrieval model and an information recall algorithm, and acquire user historical operation behavior data, and based on the vector retrieval model and the user historical operation behavior data, the recall pool configuration module calls the information recall algorithm to obtain a recall information set.

[0085] Specifically, the information recall algorithm is: the recall pool configuration module sets a similarity reference value, and based on the vector retrieval model, the user historical operation behavior data, and the similarity reference value, performs a similarity reference recall step to obtain first recall information; the recall pool configuration module sets a weight reference value, and based on the user historical operation behavior data and the weight reference value, performs a weight value reference recall step to obtain second recall information; the recall pool configuration module sets a click frequency threshold and a probability reference value, and based on the user historical operation behavior data, the click frequency threshold, and the probability reference value, performs a probability value reference recall step to obtain third recall information; and the recall pool configuration module packages the first recall information, the second recall information, and the third recall information to obtain the recall information set.

[0086] Specifically, the similarity reference recall step includes: the recall pool configuration module identifies user interest information based on the user historical operation behavior data; the recall pool configuration module screens a first interest word vector matching the user interest information from a plurality of pre-training information word vectors; the recall pool configuration module confirms a first similar vector with a vector similarity to the first interest word vector reaching the similarity reference value based on the vector retrieval model; and the recall pool configuration module intercepts first similar information matching the first similar vector in the platform information library to obtain the first recall information.

[0087] Specifically, the weight value reference recall step includes: the recall pool configuration module calculates user preference data based on the user historical operation behavior data; the recall pool configuration module screens preference information matching the user preference data in the platform information library; the recall pool configuration module calculates a preference weight of the preference information based on the user preference data; the recall pool configuration module identifies information quality of the preference information; the recall pool configuration module calculates an information weight of the preference information based on the preference weight and the information quality; and the recall pool configuration module intercepts the preference information corresponding to the information weight reaching the weight reference value as the second recall information.

[0088] Specifically, the probability value reference recall step includes: the recall pool configuration module identifies user information click data based on the user historical operation behavior data; the recall pool configuration module identifies high-frequency click information reaching the click frequency threshold in the user historical operation behavior data based on the user information click data; the recall pool configuration module predicts a secondary click probability of the high-frequency click information based on Poisson distribution; and the recall pool configuration module intercepts the high-frequency click information corresponding to the secondary click probability reaching the probability reference value as the third recall information.

[0089] In the information flow recommendation system based on the pre-training algorithm, the sorting processing module is configured to configure a first feature acquisition algorithm, a second feature acquisition algorithm, a similarity analysis algorithm, a deep learning model, and a fine sorting model. The sorting processing module performs information sorting and recommendation operations based on the pre-training word vector set, the recall information set, the first feature acquisition algorithm, the second feature acquisition algorithm, the similarity analysis algorithm, the deep learning model, and the fine sorting model.

[0090] Specifically, the information ranking recommendation operation includes: the sorting processing module calling the first feature acquisition algorithm to obtain user coarse ranking features and information coarse ranking features; the sorting processing module calling the similarity analysis algorithm based on the pre-trained word vector set, the recall information set, the user coarse ranking features, the information coarse ranking features and the deep learning model to obtain a coarse ranking information set; the sorting processing module calling the second feature acquisition algorithm to obtain fine ranking features; the sorting processing module training the fine ranking model based on the fine ranking features to obtain a to-be-used ranking model; the sorting processing module calling the to-be-used ranking model to perform fine ranking and screening processing on the coarse ranking information set to obtain a fine ranking information set; and the sorting processing module outputting the fine ranking information set to a user terminal.

[0091] Specifically, the similarity analysis algorithm includes: the sorting processing module setting a first similarity index; the sorting processing module importing the user coarse ranking features into an embedding layer of the deep learning model to obtain user coarse ranking feature word vectors; the sorting processing module accessing the pre-trained word vector set and connecting the user coarse ranking feature word vectors, the short-term pre-trained user word vectors and the long-term pre-trained user word vectors to obtain a user mixed word vector; the sorting processing module importing the information coarse ranking features into the embedding layer of the deep learning model to obtain information coarse ranking feature word vectors; the sorting processing module accessing the pre-trained word vector set and connecting the information coarse ranking feature word vectors with a plurality of pre-trained information word vectors respectively to obtain a plurality of information mixed word vectors; the sorting processing module entering a full connection layer of the deep learning model and performing similarity calculation on the user mixed word vector and the plurality of information mixed word vectors based on a cosine formula to obtain a plurality of similarity results; the sorting processing module setting the similarity result reaching the first similarity index as a qualified result; the sorting processing module setting the information mixed word vector corresponding to the qualified result as a qualified vector; and the sorting processing module intercepting recall information matching the qualified vector from the recall information set to obtain the coarse ranking information set.

[0092] Embodiment 3

[0093] The embodiment provides a computer readable storage medium, including:

[0094] The storage medium is used to store computer software instructions for implementing the information flow recommendation method based on a pre-training algorithm as described in Embodiment 1, which contains a program for executing the information flow recommendation method based on a pre-training algorithm; specifically, the executable program can be built into the information flow recommendation system based on a pre-training algorithm as described in Embodiment 2, so that the information flow recommendation system based on a pre-training algorithm can implement the information flow recommendation method based on a pre-training algorithm as described in Embodiment 1 by executing the built-in executable program.

[0095] In addition, the computer readable storage medium of the embodiment can adopt any combination of one or more readable storage media, wherein the readable storage medium includes an electrical, optical, electromagnetic, infrared or semiconductor system, device or apparatus, or any combination thereof.

[0096] Unlike the prior art, the information flow recommendation method, system and medium based on a pre-trained algorithm can realize accurate matching and recommendation of platform information based on self-developed algorithms through the method, the matching range of recommended information is low in limitation, easy to expand and extend, in the numerous information of the platform, the algorithm can be called to filter and intercept information from multiple aspects according to user demand, and the self-innovative sorting feature is added in the final fine arrangement stage, which further improves the accuracy and matching degree of the recommended information, the system provides effective technical support for the method, finally improves the user experience and the information recommendation processing efficiency, and has high application value and market value.

[0097] The above embodiment sequence number of the embodiment of the present application is only for description, not representing the advantages and disadvantages of the embodiment.

[0098] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiments can be completed by hardware, or a program for instructing related hardware to complete, and the program can be stored in a computer readable storage medium. The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk.

[0099] The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, which is made by using the content of the specification and drawings of the present application, is also included in the patent protection scope of the present application.

Claims

1. An information flow recommendation method based on a pre-trained algorithm, characterized in that, Includes the following steps: Pre-training processing steps: Set preprocessing parameters, configure pre-training algorithm and first training algorithm, and call the pre-training algorithm based on big data platform, the preprocessing parameters and the first training algorithm to obtain pre-trained word vector set; The preprocessing parameters include: a first time range, a second time range, and a third training period; The pre-training algorithm is as follows: First behavioral data of the user within the first time range is obtained based on the big data platform; the user's information click sequence is obtained based on the first behavioral data; the information click sequence is used as training data, and the first training algorithm is called to train the training data according to the third training cycle to obtain several pre-trained information word vectors; several first word vectors are identified from the several pre-trained information word vectors according to the first time range; several second word vectors are identified from the several pre-trained information word vectors according to the second time range; a weighted average calculation is performed based on the several first word vectors to obtain long-term pre-trained user word vectors; a weighted average calculation is performed based on the several second word vectors to obtain short-term pre-trained user word vectors; the long-term pre-trained user word vectors, the short-term pre-trained user word vectors, and the several pre-trained information word vectors are packaged to obtain the pre-trained word vector set. Recall pool configuration steps: Configure a vector retrieval model and an information retrieval algorithm, obtain user historical operation behavior data, and call the information retrieval algorithm based on the vector retrieval model and the user historical operation behavior data to obtain a retrieval information set; The information recall algorithm is as follows: A similarity benchmark value is set; a similarity benchmark recall step is performed based on the vector retrieval model, the user's historical operation behavior data, and the similarity benchmark value to obtain first recall information; a weight benchmark value is set; a weight benchmark recall step is performed based on the user's historical operation behavior data and the weight benchmark value to obtain second recall information; a click count threshold and a probability benchmark value are set; a probability benchmark recall step is performed based on the user's historical operation behavior data, the click count threshold, and the probability benchmark value to obtain third recall information; the first recall information, the second recall information, and the third recall information are packaged to obtain the recall information set. Sorting steps: Configure a first feature acquisition algorithm, a second feature acquisition algorithm, a similarity analysis algorithm, a deep learning model, and a fine ranking model, and perform information ranking and recommendation operations based on the pre-trained word vector set, the recall information set, the first feature acquisition algorithm, the second feature acquisition algorithm, the similarity analysis algorithm, the deep learning model, and the fine ranking model; The information ranking and recommendation operation includes: calling the first feature acquisition algorithm to obtain user coarse ranking features and information coarse ranking features; and calling the second feature acquisition algorithm to obtain fine ranking features.

2. The information flow recommendation method based on a pre-trained algorithm according to claim 1, characterized in that: The similarity-based recall steps include: User interest information is identified based on the user's historical operation behavior data; a first interest word vector matching the user's interest information is selected from several pre-trained information word vectors; a first similar vector whose vector similarity to the first interest word vector reaches the similarity benchmark value is identified based on the vector retrieval model; and the first similar information matching the first similar vector is extracted from the platform information database to obtain the first recall information.

3. The information flow recommendation method based on a pre-trained algorithm according to claim 1, characterized in that: The weighted value benchmark recall step includes: User preference data is calculated based on the user's historical operation behavior data; preference information matching the user preference data is filtered out from the platform information database; preference weight of the preference information is calculated based on the user preference data; information quality of the preference information is identified; information weight of the preference information is calculated based on the preference weight and the information quality; preference information corresponding to the information weight that reaches the weight benchmark value is extracted as the second recall information.

4. The information flow recommendation method based on a pre-trained algorithm according to claim 1, characterized in that: The probability value-based recall step includes: Based on the user's historical operation behavior data, user information click data is identified. Based on the user information click data, high-frequency click information that reaches the click count threshold is identified in the user's historical operation behavior data. The secondary click probability of the high-frequency click information is predicted based on the Poisson distribution. The high-frequency click information corresponding to the secondary click probability that reaches the probability benchmark value is extracted as the third recall information.

5. The information flow recommendation method based on a pre-trained algorithm according to claim 4, characterized in that: The information sorting and recommendation operation also includes; Based on the pre-trained word vector set, the recall information set, the user coarse ranking features, the information coarse ranking features, and the deep learning model, the similarity analysis algorithm is invoked to obtain the coarse ranking information set; The fine-ranking model is trained based on the fine-ranking features to obtain a ranking model to be used; the ranking model to be used is called to perform fine-ranking filtering on the coarse-ranking information set to obtain a fine-ranking information set; The refined ranking information set is output to the user terminal.

6. The information flow recommendation method based on a pre-trained algorithm according to claim 5, characterized in that: The similarity analysis algorithm is as follows: Set the first similarity metric; The user coarse-ranked features are imported into the embedding layer of the deep learning model for calculation to obtain the user coarse-ranked feature word vectors; the pre-trained word vector set is accessed, and the user coarse-ranked feature word vectors, the short-term pre-trained user word vectors, and the long-term pre-trained user word vectors are concatenated to obtain the user hybrid word vectors; The coarse-ranked information features are imported into the embedding layer of the deep learning model for calculation to obtain coarse-ranked information feature word vectors; the pre-trained word vector set is accessed, and the coarse-ranked information feature word vectors are concatenated with several pre-trained information word vectors to obtain several mixed information word vectors. Entering the fully connected layer of the deep learning model, the similarity between the user-mixed word vector and several information-mixed word vectors is calculated based on the cosine formula, resulting in several similarity results; The similarity result that meets the first similarity index is defined as the target result; the information hybrid word vector corresponding to the target result is defined as the target vector. The recall information that matches the target vector is extracted from the recall information set and packaged to obtain the coarse ranking information set.

7. An information flow recommendation system based on a pre-trained algorithm, based on any one of claims 1 to 6, characterized in that: include: The pre-training processing module, the recall pool configuration module, and the sorting processing module; The pre-training processing module is used to set pre-processing parameters and configure the pre-training algorithm and the first training algorithm. The pre-training processing module calls the pre-training algorithm based on the big data platform, the pre-processing parameters and the first training algorithm to obtain a pre-trained word vector set. The recall pool configuration module is used to configure the vector retrieval model and the information recall algorithm, and to obtain user historical operation behavior data. The recall pool configuration module calls the information recall algorithm based on the vector retrieval model and the user historical operation behavior data to obtain the recall information set. The ranking processing module is used to configure a first feature acquisition algorithm, a second feature acquisition algorithm, a similarity analysis algorithm, a deep learning model, and a fine ranking model. The ranking processing module performs information ranking and recommendation operations based on the pre-trained word vector set, the recall information set, the first feature acquisition algorithm, the second feature acquisition algorithm, the similarity analysis algorithm, the deep learning model, and the fine ranking model.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the information flow recommendation method based on a pre-trained algorithm as described in any one of claims 1 to 6.

Images